Download Efficient On-Chip Pipelined Streaming Computations on Scalable Manycore

Survey
yes no Was this document useful for you?
   Thank you for your participation!

* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project

Document related concepts
no text concepts found
Transcript
Publish in ACACES 2012 Poster Abstracts (HiPEAC)
ISBN 978 90 382 1987 5
For personal use only
Efficient On-Chip
Pipelined Streaming
Computations on
Scalable Manycore
Architectures
Nicolas Melot∗,1 , Christoph Kessler∗,1 ,
Jörg Keller†,2
∗
†
Linköpings Universitet, Dept. of Computer and Inf. Science, Linköping, Sweden
FernUniversitat in Hagen, Fac. of Math. and Computer Science, Hagen, Germany
ABSTRACT
Performance of manycore processors is limited by programs’ use of off-chip main memory. Streaming
computation organized in a pipeline limits accesses to main memory to tasks at boundaries of the pipeline
to read or write to main memory. The Single Chip Cloud computer (SCC) offers 48 cores linked by a highspeed on-chip network, and allows the implementation of such on-chip pipelined technique. We assess the
performance and constraints provided by the SCC and investigate on on-chip pipelined mergesort as a
case study for streaming computations. We found that our on-chip pipelined mergesort yields significant
speedup over classic parallel mergesort on SCC. The technique should bring improvement in power
consumption and should be portable to other manycore, network-on-chip architectures such as Tilera’s
processors.
1
Introduction
Scaling up the number of cores integrated in a single chip is made difficult by the mechanisms necessary to implement shared memory and keep high performance thanks to caches. In a shared memory
system, all caches maintain a coherent view of a shared memory, which requires complex hardware
that is hard to extend to serve many cores. The Single Chip Cloud computer (SCC) issued from Intel’s Tera-scale research program eliminates the hardware overhead required to maintain a consistent
view of shared memory. Instead, consistent view of shared memory in caches is maintained through
software protocols with the help of a non-uniform, low-latency shared on-die memory accessible by
all 48 cores through a high-speed on-chip network.
Multicore and manycore processors can reduce energy consumption while providing better computation power, but their parallel nature makes them very challenging to program efficiently. Many
1 E-mail:
2 E-mail:
{nicolas.melot, christoph.kessler}@liu.se
[email protected]
contributions address automatic or compiler-assisted parallelization of sequential programs, scheduling or synchronization issues for example. Some other contributions extend or create programming
languages and provide constructs to help programmers to express parallelism, for example OpenMP
or Streamit [TKA02]. Streamit organizes computation in a succession of filters linked by channels
from which filters can read input data and write the corresponding output. Depending on their characteristics, they can be mapped to different cores or accelerators in order to achieve task or pipeline
parallelism [UGT09].
The heavy use of main memory impacts both speed and energy performance of multi and manycore processors [HKK10]. On-chip pipelining [HKK10] reorganizes computations into a task graph
where tasks can be compared to Streamit filters. Tasks are mapped to cores and forward intermediate
results to the next tasks in the graph through on-chip network and distributed memory, or via cores’
local memory if mapped to the same core. This reduces accesses to main memory strictly to nodes
having to fetch input data or write final computation results, and thus frees intermediate communications from costly latencies thanks to the high-speed on-chip network. Such property would benefit to
software running on MPSoC, whose design is heavily constrained by energy consumption issues and
that already makes heavy use of streaming computations; for instance to process image sequences or
Fast Fourier Transform.
We use mergesort as a simple streaming computation case study to demonstrate the usefulness of
the on-chip pipelined approach to parallel computation. We investigated performance and limitation of
the SCC and we implemented a non-pipelined parallel version of mergesort as a comparison basis to
the on-chip pipelined variant. We introduce an ILP (Integer Linear Programming) formulation to solve
the task-to-processor mapping problem, taking into account the constraints measured in previous
work. Finally, we implemented on-chip pipelined mergesort and compare its performance with nonpipelined parallel mergesort for SCC.
The rest of the paper is structured as follow: section 2 introduces the SCC architecture and the
performance constraints we observed. Section 3 introduces the on-chip pipelining approach and its
implementation on the SCC. Finally, section 4 concludes and introduces future work.
2
The Single Chip Cloud computer
The SCC provides 48 independent Intel x86 cores, organized in 24 tiles. Figure 1(a) provides a global
schematic view of the chip. Tiles are linked together through a 6 × 4 mesh on-chip network. Each
tile embeds two cores as well as a common message passing buffer (MPB) of 16KiB (8KiB for each
core); the MPB supports direct core-to-core communication.
The cores are IA-32 x86 (P54C) cores which are provided with individual L1 and L2 caches of
size 32KiB (16KiB code + 16KiB data) and 256KiB, respectively, but no SIMD instructions. Each
link of the mesh network is 16 bytes wide and exhibits a 4 cycles crossing latency, including the
routing activity.
Each core is attributed a private domain in main memory whose size depends on the total memory
available (max 64 GiB). Six tiles (12 cores) share one of the four memory controllers to access cores’
private memory. Cores can access an off-chip shared memory whose size can vary up to several
hundred megabytes. Note that private memory is cached on cores’ L2 cache but caching for shared
memory is disabled or offers no coherency among cores’ caches. Intel provides the RCCE library
which contains MPI-like routines to synchronize cores and allows them to communicate data to each
other. RCCE also allows the management of voltage and frequency scaling.
Previous work assessing the SCC shows a significant performance penalty from the use of shared
main memory, compared to using private memory or the on-chip network [AMKK11]. It also indicates that the distance to the memory controller has an influence on round-trip time. Further tests
indicate the difficulty to actually saturate the memory controllers using all cores to run read and
write operations. We also see that there is more read that write memory bandwidth available to the
cores [MAKK11]. Finally, we observe memory bandwidth drops with cache-unfriendly memory access patterns and even more drops with random access patterns [MAKK11].
23
15
3
7
22
0
1
14
2
RAM
RAM
21
5
13
6
12
10
9
8
19
18
17
16
RAM
20
4
11
(a) A schematic view of the SCC die.
(b) A six-level merging tree mapped to 6 cores.
Figure 1: (left) A schematic representation of the SCC die. (right) An on-chip pipelined task graph.
3
On-chip pipelined mergesort
On-chip pipelined mergesort organizes mergesort’s merging tree in a pipeline, with all tasks running
concurrently on an arbitrary number of cores (see Fig.1(b) for a 6 cores example). Producer and
consumer tasks mapped to the same core communicate through its local memory; those mapped to
different cores use the mesh network and distributed on-chip memory. The leaf tasks start the computation by reading data from memory and they forward their intermediate merging results towards
their consumer tasks at the next level in the task graph. These tasks can then start the same process,
and so on, until all the tasks in the pipeline are active. This schema restricts memory writes to the root
node. As the merge operation is done blockwise, follow-up tasks can start as soon as leaf tasks have
produced their first block of intermediate results.
On-chip pipelined mergesort implemented on SCC follows the same process. Because of the high
amount of cores available (48), the chip is divided into 8 groups of 6 cores each, all of them using
the same memory controller but located on a different tile. This restricts the task mapping problem
to a 6-level merging tree placed on 6 cores, simplifying significantly the work of the ILP solver
to compute an optimal task mapping [MKA+ 12]. The symmetry of the SCC allows 8 replicas of
this merging tree with no or marginal performance losses. The mergesort implementation runs 8
pipelined mergesort and merges the 8 sorted subsequences using a classic parallel mergesort and
shared memory [MKA+ 12].
The on-chip pipelined phase of our implementation yields a speedup between 1.6 and 1.7, depending on the adaptations of the classic parallel mergesort we compare to [MKA+ 12]. Furthermore,
we observed a speedup of 143% from a previously implemented on-chip pipelined mergesort over
CellSort for the Cell B.E. [HKK10].
4
Conclusion
Although multi and manycore processors improve computation power and energy consumption over
sequential processors, their performance is limited by main memory’s latencies. Organizing computation in a pipeline not only enables pipeline parallelism, but it can also reduce off-chip memory
accesses, yielding better speedup. Our experiments, although still using suboptimal task mappings,
show that this technique can provide good runtime improvements over classic parallel implementations. We believe these results can be further improved with a refined implementation and the use of
optimal, ILP-generated task mapping. On-going work includes porting the on-chip pipelining technique to other calculations such as Fast Fourier Transform or other streamed computations. We expect
the reduced usage of off-chip memory also yields better energy consumption. Furthermore, reduced
and easy to control accesses to main memory yields determinism that enables the use of on-chip
pipelining in high-performance real-time systems. The technique should be portable to other manycore, network-on-chip architectures such as Tilera’s processors or Adapteva’s Epiphany architecture.
Acknowledgments
The authors are thankful to Intel for providing the opportunity to experiment with the “conceptvehicle” many-core processor “Single-Chip Cloud Computer”.
This research is partly funded by the Swedish Research Council (Vetenskapsrådet), project Integrated Software Pipelining, and the CUGS graduate school at Linköping University.
References
[AMKK11] K. Avdic, N. Melot, C. Kessler, and J. Keller. Parallel sorting on Intel Single-Chip
Cloud Computer. In Proc. A4MMC workshop on applications for multi- and many-core
processors at ISCA-2011, 2011.
[HKK10] R. Hultén, J. Keller, and C. Kessler. Optimized on-chip-pipelined mergesort on the Cell/B.E. In Proceedings of Euro-Par 2010, volume 6272, pages 187–198, 2010.
[MAKK11] N. Melot, K. Avdic, C. Kessler, and J. Keller. Investigation of main memory bandwidth
on Intel Single-Chip Cloud Computer. Intel MARC3 Symposium 2011, Ettlingen, 2011.
[MKA+ 12] N. Melot, C. Kessler, K. Avdic, P. Cichowski, and J. Keller. Engineering parallel sorting
for the Intel SCC. In Proceedings of 4th Workshop on using Emerging Parallel Architectures (WEPA 2012), 2012.
[TKA02] W. Thies, M. Karczmarek, and S. Amarasinghe. Streamit: A language for streaming
applications. In Compiler Construction, volume 2304 of Lecture Notes in Computer
Science, pages 49–84. 2002.
[UGT09] A. Udupa, R. Govindarajan, and M.J. Thazhuthaveetil. Synergistic execution of stream
programs on multicores with accelerators. In Proceedings of ACM SIGPLAN/SIGBED
2009 Conference on Languages, Compilers, and Tools for Embedded Systems (LCTES
2009), pages 99–108, 2009.